Attention-Based Noise Prior Network for Magnetic Resonance Image Denoising

Data obtained from magnitude magnetic resonance images (MRI) are corrupted by signal-dependent and spatially variant Rician noise. The contribution of this work can be discussed at three different levels. Firstly, Rician noise-level estimation map is fed as prior concatenated with the noisy input data; to handle spatially variant noise and achieve an optimal compromise between noise removal and detail preservation. Secondly, modified U-Net architecture is used to accommodate non-local multi-level and multi-scale features. Thirdly, to preserve the long-range dependencies lost in farther symmetric layers, features obtained from symmetric-group attention block is fed as input to the deconvolution layers for nonlocal high- and low-level feature mapping. Experimental results over synthetically corrupted MR images and real data obtained from MR scanners suggest the potential utility of our proposed technique over a wide-range of noise levels.